On Wed, 16 Sep 2026, Uros Bizjak wrote: > Set the i386 default for MD generator partitions to 6:6:1. > > The longest objects in the bootstrap are gimple-match-[1..10].o, > with per-file user times of: > > user 0m12,114s > user 0m6,214s > user 0m6,737s > user 0m8,234s > user 0m5,765s > user 0m5,221s > user 0m5,732s > user 0m5,221s > user 0m5,645s > user 0m7,951s > > while generic-match-[1..10].o compile in: > > user 0m5,875s > user 0m5,186s > user 0m4,132s > user 0m4,367s > user 0m4,464s > user 0m4,705s > user 0m4,340s > user 0m5,942s > user 0m4,469s > user 0m4,236s > > With the default set to 6:6:1, compile times become: > > insn-emit-[1..6].o: > > user 0m4,766s > user 0m4,799s > user 0m4,829s > user 0m4,825s > user 0m4,766s > user 0m4,781s > > insn-recog-[1..6].o: > > user 0m4,782s > user 0m4,741s > user 0m4,988s > user 0m4,814s > user 0m4,918s > user 0m4,898s > > insn-attrtab-1.o: > > user 0m3,076s > > which are all comparable to the generic-match-[1..10].o compile times. > > Raising the partition counts further would shorten objects that are > already off the critical path, while each additional partition > repeats the per-file preamble, adding to overall bootstrap CPU time. > > gcc/ChangeLog: > > * config.gcc (i[34567]86-*-* | x86_64-*-*): Set > default_md_generator_partitions target default to 6:6:1. > > Bootstrapped on x86_64-pc-linux-gnu.
LGTM > BTW: It looks like 6:6:1 should be the default for all targets. i386 > has one of the largest insn-*.cc sources and measurements confirm that > aarch64 default is also optimal for i386. The question is where the default of 10 came from? But sure. Richard. > Uros. > -- Richard Biener <[email protected]> SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg, Germany; GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)
